Candid accounts of building, closing and rebuilding companies. Decisions, mistakes and changing beliefs, alongside essays on ownership and technological change.
Startups in San Francisco raise $18 million just to afford the compute. We did the same work bootstrapped. The real gate on who gets to build AI isn't ability - it's who can afford to be in the room.
For six months I was sure the answer was a from-scratch Canadian AI model. Seven days ago that broke. I had confused the model with the system - and the open-book version is the bigger bet.
Yesterday I published the post-mortem: we asked flash-1-mini ten questions any Canadian lawyer would consider basic, and it invented seven citations. That post was about what broke. This one is about what it changed. Because two weeks ago I was telling people we were building a Canadian legal AI model - and today we decided we're not. I want to walk through why, because the answer changed how I think about what "building AI" actually means. The original scope The original scope made sense
An honest post-mortem on our own model — and what anyone building or buying AI should take from it.
Yesterday we released CBLRE — the Canadian Bilingual Legal and Regulatory Evaluation. The day before, we released flash-1-mini, a 4-billion-parameter bilingual Canadian legal AI model. Most of the launch coverage has focused on the model. That's the wrong artifact to focus on. The model is the proof. CBLRE is the moat. Here's why. The gap nobody had filled Before yesterday, no standard public benchmark existed for Canadian bilingual legal AI evaluation. That sentence is bigger than it so
Two AI infrastructure announcements landed in the same week. Ours was a model. We released flash-1-mini — a 4-billion-parameter bilingual Canadian legal AI — under Apache 2.0, alongside an open benchmark (CBLRE), an open training corpus, and the methodology behind both. The model runs on a MacBook. It runs on a Raspberry Pi. The weights are downloadable. You own the file. Perplexity's was an orchestrator. At Intel's Computex keynote in Taipei, CEO Aravind Srinivas demonstrated what the com
The model is the smallest part of the story. Here's what it is, what it isn't, and what comes next. Today we shipped flash-1-mini. It's a 4-billion-parameter open-weight model, fine-tuned for Canadian context, bilingual in English and French, that runs on a laptop with no cloud dependency. You can download it, run it offline, and own it. The weights are yours. I want to write about what it is, what it isn't, and what comes after — because the model itself is the smallest part of the story.
Over the past two years, Canada quietly rebuilt the question of who controls its artificial intelligence. Not in one announcement. That's the point. There was never a single moment loud enough to make you look up. The decisions arrived in fragments — a Christmas Eve letter, a contribution agreement with a file number, a press release at a university most people don't follow, an MOU with the important parts blacked out. Each fragment was, on its own, small enough to ignore. Together they red
Two A100 graphics cards. Spinning quietly in a Google datacenter. Five hours of training. About $50 in compute. That's what it cost us to fine-tune our own 4-billion-parameter AI model this week. The base model went from 30% accuracy on the tasks we care about to 98%. Read any article on fine-tuning costs and you'll see numbers between $5,000 and $35,000. One blog called it a 'CFO conversation.' Another listed 'hidden expenses' that could double your initial estimate. A third quoted team
Three weeks ago I wrote a post called GPU Cloud Shopping in Canada: What's Actually Available. The short version: I checked every major cloud provider with a Canadian data center, trying to rent a current-generation GPU to train AI models in this country. Google Cloud Montreal had chips from 2017. AWS listed the right hardware but wouldn't let me actually run it. OVHcloud's H100s turned out to be in France, not Quebec. DigitalOc
Training an AI model is assumed to cost millions of dollars. It's the single most common misconception in the space, and it's wrong by roughly two orders of magnitude for the activity most people actually want to do. This post is a short, concrete breakdown of what fine-tuning actually costs in 2026, what it doesn't cost, and where the real spend lives. I'm writing it now because 'how much does this cost' is the first question
I'm training an AI model. It's going to run on a laptop. Three weeks ago I would have told you I was training a 70-billion-parameter model, the kind of thing that needs a data center to breathe. I'm not. I'm training a 4-billion-parameter model that runs on a Mac Mini. If the smaller one works, a larger companion model may follow. But the 4B is the bet. This is the first post in a series where I'll share what I'm building, why,
Showing 12 of 130
Open next page